DramaChain Bench: An End-to-End Benchmark for Short-Drama Generation

1 Introduction

• Prior benchmarks leave pipeline handoffs unevaluated, use isolated stages and clean or authored inputs, and lack localized defect supervision.Consequently, they cannot identify which stage caused a delivery defect, how far it propagated, or which dimension failed and where.Source: §1 · p. 3
• DramaChain Bench scores every stage on pipeline-produced items across five axes and six granularities, resolving into 63 leaf dimensions over 5,785 items.The axes are input fidelity, internal consistency, generation plausibility, visual quality, and cinematic expressiveness.Source: §1 · p. 3
• DramaChain Agent generates the benchmark’s 20 dramas and 60 episodes, validated against production platforms, and forks comparisons at the tested stage with identical upstream artefacts and prompts.All other stages remain fixed, preventing a flawed pipeline from becoming the object of evaluation.Source: §1 · p. 3
• Three professional annotators independently score items using five-point leaf-dimension rubrics, producing 17,488 scores and 255,925 reviewable attributions; Agentic Judge gathers evidence across rounds before checklist-based judgments.The annotations localize deductions spatio-temporally and map rubric criteria to a closed vocabulary of attribution tags.Source: §1 · p. 3
• 3.30: gpt-5.5-xhigh, gpt-image-2, and seedance-2.0 lead their stages and deliver a short drama above the 3.0 usable line.Degrading one upstream stage costs a clip 0.07–0.13 points but the finished short drama 0.25–0.83; Agentic Judge reproduces the model board at mean PLCC 0.918.Source: §1 · p. 4
• Stage-level branching and model-consistent automatic scoring allow new models to be evaluated using single-stage generation without additional human annotation.The authors will release DramaChain Dimensions, the Agentic Judge framework, and a curated benchmark-data subset.Source: §1 · p. 4

2.1 Storytelling video generation

• Storytelling-video methods span joint multi-shot generation and staged agent pipelines, while long-range consistency degrades and available controls remain too coarse for stage-specific correction.Existing controls include prompts, camera paths, and storyboard sketches; incorrect requirements cannot be re-issued to the responsible stage without rerunning the whole pipeline.Source: §2.1 · p. 5
• Short-drama pipelines expose hand-off challenges: text storyboards underspecify renderer geometry, and errors can accumulate across script, keyframe, and video stages.Commercial platforms use frameworks of the same general shape, motivating evaluation at each stage.Source: §2.1 · p. 5

2.2 Datasets and benchmarks

• Existing corpora lack both a full episode chain produced by one pipeline and annotations that localize defects and attribute them to a stage.Most use human-made inputs for the stage under test and clip-level ratings or captions rather than defect locations and causes.Source: §2.2 · p. 5
• Prior automated judges are reported as competitive on semantic alignment but weak when temporal resolution matters.The paper says its later per-dimension results reproduce this split.Source: §2.2 · p. 5
• Benchmark research broadens quality dimensions and targets, but workflow-aware approaches still do not score intermediate artefacts against stage-level rubrics.Some evaluate workflow completion or align criteria while evaluating finished video; short-drama suites also stop short of the full chain.Source: §2.2 · pp. 5–6

3.1 DramaChain Dimensions: five axes across six stages, 63 leaf dimensions

• Five evaluation axes are shared across six production stages and resolve into 63 leaf dimensions, making defects traceable from upstream instructions to downstream outputs.The axes are input fidelity, internal consistency, generation plausibility, visual quality, and cinematic expressiveness; each leaf dimension has a five-level description and fixed attribution tags.Source: §3.1 · p. 7, Figure 2 · p. 7

3 DramaChain Bench

• The benchmark pipeline produces its own items and intermediate production artefacts, with its chain reverse-engineered from commercial platforms and optimised.It is built on the ViMax framework [17] and produces character portrait sheets, reference sheets, per-shot descriptions, and a camera tree alongside the six scored stages.Source: §3 · p. 6, §3.2 · p. 8

3.2 DramaChain Agent: building a short-drama production system to commercial-platform parity

• DramaChain Agent generates 60 episodes across 20 dramas and forks only at the tested stage, giving competing models identical upstream artefacts and prompts.The pipeline was benchmarked against three commercial platforms end to end without human editing or best-of-n reruns; expert review found quality in the same band, with stylistic differences.Source: §3.2 · pp. 8–9

3.3 DramaChain Labeling System: treating the reference answer as an engineering artefact

• Three annotators independently scored each item, producing 17,488 scored entries and 255,925 verifiable attributions that localise defects in space and time.The median annotation time was 23 minutes per item; reviewers sample team outputs, failed annotations are reassigned, and a dashboard flags low-quality submissions using nine heuristic rules.Source: §3.3 · pp. 9–10

3.4 DramaChain Agentic Judge: multi-round agentic scoring

• Human annotations show most specialist metrics correlate weakly with judgments (|SRCC| ≤ 0.2), including negative correlations for fourteen photography-oriented metrics.AIGC-specific perceptual models also remain below 0.2; the authors attribute this to divergent priors, and identify three leaf dimensions, including V-D2, with near-random human baselines.Source: §3.4 · p. 11
• DramaChain Agentic Judge scores each leaf dimension independently over multiple rounds, using relevant measurements as auxiliary cues and item-specific checklists for verdicts.Agents can inspect regional zooms and timestamped frames, and return defect tags and spatio-temporal localisations for cross-checking disagreements with human annotations.Source: §3 · p. 7, §3.4 · p. 11

4.2 Main results

• Among annotated models, the strongest chain—gpt-5.5-xhigh for storyboarding, gpt-image-2 for keyframes, and seedance-2.0 for video—barely meets the usable line: its finished drama scores 3.30 (0.3 above), and its best single-shot result is 3.24.Source: §4.2 · p. 11
• Evaluation gaps are widest on reference-anchored criteria, while similar composite scores can conceal large, opposing dimension-level strengths.Automated top-to-bottom gaps are 1.22 for F input fidelity, 0.70 for C internal consistency, 0.64 for Q visual quality, 0.53 for P generation plausibility, and 0.42 for E cinematic expressiveness. For example, seedream-5.0-pro and nano-banana-pro differ by 0.02 overall but by +0.39 on F and −0.60 on P, illustrating why composites support tiering better than model selection.Source: §4.2 · p. 14

4.3 Agreement validation

• The Judge reproduces model rankings well overall, but item-level agreement varies by stage and depends on modality and task characteristics.It achieves mean PLCC 0.918 across six stages and SRCC 1.000 for short dramas; text and image stages meet or exceed the human baseline, while video reaches 0.79–0.85×. Agreement declines from text (0.509) to images (0.440) to video (0.296); enumerability and fine-grained perception, rather than subjectivity, predict automatability, and only F improves over the human baseline overall.Source: §4.3 · pp. 15–16, Table 3 · p. 15, Table 4 · p. 16

4.4.1 Storyboard design

• Storyboard models largely preserve core script elements, but production logic—especially shooting rhythm—is the text stage’s main bottleneck.Seven of eight leaf dimensions meet the usable line, while shooting rhythm scores 2.74 and restraint in additions 3.34. The dominant defect is duration budgeting, which can be checked before visual rendering; validated automation also scores seven additional models on the same items at zero incremental annotation cost, with intra-family gains of +0.047 to +0.582.Source: §4.4.1 · pp. 17–18, Table 5 · p. 17

4.4.2 Keyframe images

• Keyframe failures center on adherence to specified content and cross-shot continuity, not basic image rendering.Single-panel scores below 3.0 concern subject attributes (2.75), scene reference (2.72), and camera framing or style (2.91); at episode level, scene consistency (2.50) falls below character consistency (2.87). Clear references support stronger automated agreement, while several physically objective quality criteria remain below human baselines when the item does not specify the intended answer.Source: §4.4.2 · pp. 18–19, Table 6 · p. 19

4.4.3 Shot video and short drama

• Video models struggle most with temporal behavior, and cross-shot consistency scores worsen when evaluation shifts from clips to episodes.Audio-visual sync scores 1.49, the stage’s lowest leaf score, while style consistency scores 4.15; all four shared cross-shot consistency dimensions decline at episode granularity. Script-verifiable event completion reaches 0.363 agreement against a 0.200 baseline, whereas watchability of the cut reaches 0.074 against 0.343.Source: §4.4.3 · pp. 19, 21, Table 7 · p. 20
• Single-shot video paradigms trade strengths according to which visual constraints their inputs provide.The grid-panel paradigm leads in subject–scene integration (4.28), character appearance consistency (3.67), and style consistency (4.32); first/last frames lead in keyframe adherence (4.28), while multi-reference leads in camera movement plausibility (4.09) and sound and music (3.62).Source: §4.4.3 · p. 21, Table 8 · p. 22

4.5 Chain-level analysis

• Upstream substitutions and coarser evaluation reveal losses hidden by isolated assets, as inherited defects accumulate across outputs.Replacing one upstream model reduces scores by 0.07–0.13 for individual clips and by 0.25–0.83 for multi-episode outputs. In one example, individually scored shots range from 3.23 to 4.00 but score 1.67 as a continuous segment because of cross-shot mismatches, showing that single-asset scores are not a valid proxy for episode quality.Source: §4.5 · pp. 22–23, Figure 10 · p. 22, Figure 11 · p. 23
• Agreement declines along the pipeline largely because of modality: input fidelity performs strongest within every modality, while every axis except P degrades as evidence adds pixels, temporal dynamics, and audio.The effects are additive. For cinematic expressiveness, all four dimensions reach the human baseline for storyboard-script requirements, versus zero of eleven when identical requirements are rendered into video; the proposed remedy is better perception, not prompt engineering.Source: §4.5 · p. 23, Table 9 · p. 23
• Running the pipeline and forking only at the stage under test attributes defects to the stage that produced them, rather than to the last stage that touched them.For example, a storyboard that compresses three speakers into one shot is recorded against storyboard design, while keyframe-induced scene drift is recorded against the keyframe stage; this attribution property underlies the section’s results.Source: §4.5 · pp. 23–24

5 Conclusion

• DramaChain Bench evaluates each production stage on in-chain outputs, using calibrated generation, localized human scoring, and a human-validated agentic judge to attribute defects to their originating stage.DramaChain Agent generates items and forks only at the stage under test; annotators localize deductions in space and time using a fixed vocabulary.Source: §5 · p. 24
• The pipeline, dimension system, and scoring configurations are reusable, and new models can be evaluated at any individual stage without additional human annotation.Human-dependent components are explicitly exposed rather than obfuscated.Source: §5 · p. 24
• The authors encourage diagnosis at short-drama failure points across inter-stage boundaries and throughout episode shots, rather than evaluating individual clips in isolation.They prefer DramaChain Bench as a diagnostic tool rather than a leaderboard.Source: §5 · p. 24

6 Limitations

• Table 10 explicitly states what the benchmark does not cover and why, qualifying the scope of its 63-dimension taxonomy.The authors note that the taxonomy might otherwise be assumed exhaustive.Source: §6 · p. 24, Table 10 · p. 24

• All benchmark artefacts are synthetic, use project-original scripts and model-generated assets, and involve no real-person likenesses, voices, or reused commercial-drama assets.Commercial platforms were queried through public interfaces using the researchers’ scripts, and only returned outputs are reported.Source: Ethics and Responsible Use · p. 24
• Items were screened against commercial release policies, but no broader harm audit was performed, and episodes without evaluable outputs were marked missing rather than imputed.The corpus retains genre tropes including conflict, coercion, and revenge.Source: Ethics and Responsible Use · pp. 24–25
• Professional contracted third-party vendor staff annotated items, with recruitment filtered by stage-specific background and median working time of 23 minutes per item.Compensation followed vendor terms, but hourly pay rates were unavailable because the researchers did not set them.Source: Ethics and Responsible Use · p. 25
• The benchmark supports stage-wise model comparison, defect diagnosis, and identification of automatable evaluation dimensions, but cannot predict commercial performance or evaluate safety and content-policy compliance.It does not certify any artefact as fit for public release.Source: Ethics and Responsible Use · p. 25

A The Dimension System and Its Implementation

• 66 leaf dimensions are defined, but 63 are scored after retiring three with near-zero human baselines; each has a five-level rubric and attribution-tag vocabulary shared by humans and the automated scorer.Per-dimension scores and agreement are reported in the per-stage tables of Sec. 4.4.Source: A The Dimension System and Its… · p. 32

A.1 Implementation of every leaf dimension

• Each scored leaf dimension uses a five-level rubric and closed tag vocabulary, while a declarative method line routes evidence producers and guides further tool calls.The scorer emits a band, tag, and reason; pixel-sensitive dimensions cannot use pixel-altering tools, preventing image assessment on retouched frames. Some routed evidence is unavailable or withheld for specific stage limitations.Source: §A.1 · p. 32

1 ⃝Storyboard design

• Storyboard dimensions assess event coverage and dialogue fidelity, description executability and restraint, narrative flow, and shooting rhythm through alignment, language-model, and visual-model checks.The listed methods include beat-to-shot coverage alignment, dialogue text alignment, per-shot checks for surplus detail and gaps, and sequence and duration assessments.Source: §1 · pp. 32–33

2 ⃝Single keyframe

• Single-keyframe checks combine zoom-based defect inspection, shot framing and style assessment, face-identity similarity, and zoomed comparison of appearance and costume against character sheets.Face identity uses ArcFace for live action or an anime-ID model or DINOv3 feature similarity with CIDS [43] [78], selected by the audit.Source: §2 · p. 33

3 ⃝Episode keyframes

• Episode-keyframe checks assess cross-shot character identity, scene consistency, and style, while spatial-orientation coherence is marked retired.Character consistency aggregates pairwise cross-shot identity similarity with gap decay; scene consistency uses cross-shot scene-bank embeddings and zoomed set-dressing checks.Source: §3 · p. 34

4 ⃝Single-shot video

• Single-shot video dimensions cover perceptual image quality and motion texture, sampled-frame attributes and framing, and consistency of characters, scenes, props, style, actions, and on-screen text.Methods include perceptual quality and sharpness measures, frame sampling with SAM and detection counts, per-frame consistency checks, and OCR edit-distance comparisons across adjacent frames.Source: §4 · p. 34
• Single-shot video also evaluates motion smoothness, physical and causal plausibility, speech quality, audio-visual synchronization, and how music and sound effects fit scene emotion and action.Checks include inter-frame optical flow, VideoPhy-AutoEval, time-window causal-order and completion judgments, DNSMOS and ASR WER/CER, SyncNet and ImageBind, and scene-fit assessment.Source: §4 · pp. 34–35

5 ⃝Episode video

• Episode-video dimensions cover cross-shot character, scene and prop, style, action, subject–scene, text, causal, speech, sound, and performance continuity.Evidence mechanisms include per-shot identity comparison with gap decay, scene- and style-bank comparisons, action and pose matching, OCR text comparisons, and ordered sub-agent judgments of causality and closure.Source: §5 · p. 35

6 ⃝Short drama

• Short-drama dimensions assess scripted event completion, style, cross-episode consistency and progression, cut watchability, and pacing and transitions.Event completion is checked against script key events; progression is judged against script beats, while pacing uses human-cut distribution gaps and transition naturalness.Source: §6 · pp. 35–36

A.2 The attribution-tag vocabulary

• The attribution vocabulary spans 66 dimensions, with 298 tag slots drawn from 253 distinct tags; every score below 5 requires a declared tag and free-text reason.The same tag lists are used by annotators and automated scoring, and the corpus contains 255,925 human attributions.Source: §A.2 · p. 36

1 ⃝Storyboard design

• Storyboard-design tags cover event and dialogue fidelity, executable shot descriptions, restraint in additions, narrative flow, shooting rhythm, emotional expression, and audiovisual style.The vocabulary includes issues such as omitted or invented plot events, unclear blocking, abrupt transitions, monotonous camera work, and genre-mismatched treatment.Source: §1 · p. 36

2 ⃝Single keyframe

• Single-keyframe tags cover image quality and style, subject attributes and actions, scene layout, framing, identity and costume, panel consistency, spatial orientation, and physical plausibility.Defects include identity drift, changed spatial structure or lighting direction, anatomical errors, and objects floating or clipping through geometry.Source: §2 · pp. 36–37

3 ⃝Episode keyframes

• Episode-keyframe tags assess cross-shot character and scene consistency, subject–scene integration, spatial-orientation coherence, and style consistency.Listed defects include identity drift, unchanged composition or pose, scene-layout changes, left/right swaps, and abrupt camera-position changes.Source: §3 · p. 37

4 ⃝Single-shot video

• Single-shot-video dimensions and tags address image quality, framing, motion and action, appearance and scene consistency, causality, speech and audiovisual sync, sound, composition, camera movement, and performance.Model-callable evidence includes composition checks using shot size, depth, blocking, and subject-box head crops, plus camera-move intent and acting or expression naturalness.Source: §4 · pp. 35, 37

A.3 The scoring prompt

• A shared prompt template scores each of 66 dimensions on a five-band scale, with dimension-specific rubrics, checklists, applicability rules, tags, evidence, and JSON output requirements.Scores below 5 require a specific checkable defect and at least one dimension-specific tag. Tool-enabled dimensions gather evidence over multiple rounds, then withdraw tools for the final verdict; severity deductions reflect the defect’s overall effect rather than accumulating per blemish.Source: §A.3 · pp. 38–39

A.4 The evidence tool chain

• The tool registry separates tools the automated scorer may call from routed measurements that always run; no tool produces the rubric score.Several tools return numbers of their own.Source: Table 11 · p. 39, §A.4 · p. 39
• 14.6 minutes of median wall-clock time per item includes 198.7 seconds of median tool time per dimension, 20 sub-agent calls, and about 19 measurement calls per job.Tools execute concurrently, making cumulative machine time considerably larger; nearly all cost comes from tool execution, so additional reasoning rounds are close to free while adding another tool is not.Source: §A.4 · pp. 39–40
• Metric routing reaches 17 of 21 video leaf dimensions; ocr_track is the only primary tool below 98.9% availability, at 71.1%.Source: §A.4 · p. 39

A.5 Validating candidate metrics

• Candidate metrics are tested field by field against human consensus at each granularity, and retained small models provide low-level evidence rather than scores.Shown values are measurements with directions rather than calibrated bands, except for physics, reference-similarity, and voiceprint fields; the aesthetic scorer is never shown, and suppression is stage-specific.Source: §A.5 · p. 40
• Most candidate metrics miss the scoring-anchor threshold of |SRCC| ≥0.3 over at least 100 items, including NIMA at −0.152 and UniPercept’s image-quality head at 0.190.Other reported examples include Laplacian sharpness −0.144, TOPIQ-NR −0.067, MUSIQ −0.057, Q-Judger 0.004, ArtiMuse 0.030, and UniPercept’s aesthetic head 0.110. The shortfall spans metric families; the authors suggest genre- and art-direction-dependent criteria may contribute.Source: §A.5 · p. 40

B Drama Manifest

• The corpus comprises 20 scripts of 10 episodes each, with the video side using the first three episodes per drama: 60 episodes stratified across channel, period/theme, genre, visual style, and dialogue language.Scripts pass an automatic item check and an expert screenwriter assessment, with one draft retained per item.Source: B Drama Manifest · p. 41

C.1 Prompt parity across models

• LLM- or VLM-driven prompt rewrites occur in a shared prefix before model branches, while moderation rewrites and duration clamping are applied per branch at call time.Models with different duration caps receive separate shared prefixes rather than a common clamping range.Source: §C.1 · p. 41

C.2 Coverage gaps

• Model coverage is incomplete and nonrandom: several models run only subsets or paradigms because of content review and generation failures, while storyboard coverage is nearly complete across nine models.Models are scored on the slices they ran, and Table 2 marks affected models with ‡.Source: §C.2 · p. 41

C.3 Absolute-score accuracy and stage bias

• Automated-score bias by stage ranges from −0.35 for single-shot video to +0.57 for episode keyframes, with the scorer lenient at five stages and strict at one.The reported biases are storyboard +0.28, single keyframe +0.41, episode keyframes +0.57, single-shot video −0.35, episode video +0.14, and short drama +0.42.Source: §C.3 · p. 41

D Content Slices of the Automated Score

• Across the content slices, input paradigm produces the largest movement in automated absolute scores, visual style the second largest, and dialogue language the smallest.The slices vary generated artefact by visual style and dialogue language, and vary the input paradigm supplied to the model.Source: D Content Slices of the Automated Score · p. 41

D.1 By visual style: live action scores higher

• Live action has the higher automated composite at five of six granularities, by 0.08 to 0.22, while single keyframes favor animation by 0.08.The comparison is within models that ran both styles.Source: §D.1 · p. 41, Table 15 · p. 43
• Style changes which defects are found: at single-shot video, C is 0.32 higher on live action but Q is 0.25 higher on animation, while F is identical.V-C3 style consistency runs over a point in favor of live action, whereas V-A1 image quality and material favors animation by two thirds of a point.Source: §D.1 · pp. 41, 43

D.2 By dialogue language: small, and confined to fidelity

• The language comparison holds scene descriptions in Chinese and balances the two groups on visual style; the four English-dialogue dramas differ from the other sixteen only in dialogue and on-screen text.Table 12 describes the English-dialogue titles as using Chinese scene descriptions, isolating language from production difficulty.Source: Table 12 · p. 42, §D.2 · p. 43
• The automated composite is higher for English at every granularity but by no more than 0.12; nearly all movement is F, which is 0.64 higher on English for the short drama, while Q is identical to two decimals where measured.The dimensions that move include event completion, speech quality, and dialogue fidelity; cross-shot character and cross-episode consistency move in the other direction.Source: §D.2 · p. 43, Table 15 · p. 43

D.3 By input paradigm: largest, and it reverses along the chain

• Input paradigm has the largest average effect and reverses: first/last frame leads grid panel by 0.37 on single keyframes, then trails it by 0.42 on episode keyframes.Grid panel tops all three video granularities, narrowly at single-shot video and the short drama; on the ten shared single-keyframe dimensions, first/last-frame's lead widens to 0.61 because the C axis exists only for grid panel.Source: §D.3 · p. 43, Table 15 · p. 43
• First/last frame leads keyframe-stage F input fidelity (3.57 against 2.73) and frame-anchored single-shot dimensions; grid panel leads episode-keyframe cross-shot consistency (3.53 against 3.11), while multi-reference is last on F (2.92 against 3.39).Grid panel leads composition but costs audio-visual sync; multi-reference leads camera plausibility, audio-visual sync, and E cinematic expressiveness, while first/last frame trails on camera plausibility.Source: §D.3 · p. 44

E Agreement Slices of Automated Evaluation

• The agreement slices hold the corpus fixed while varying stage, axis, modality, and human-judgment reproducibility; agreement is reported as correlation against annotator consensus.Both boards are continuous means, so the paper uses correlation rather than weighted kappa and reports effects in score units for comparison with the 3.0 usable line.Source: E Agreement Slices of Automated… · p. 44

E.1 By stage: the baseline is crossed between images and video

• Image-stage agreement exceeds the human baseline (1.19–2.35×), while video-stage agreement falls below it (0.79–0.85×), so the crossing follows modality rather than chain depth.Episode keyframes still clear the baseline despite requiring cross-panel comparison; a single video clip does not.Source: §E.1 · p. 44

E.2 By axis: agreement falls as the criterion loses its external reference

• Agreement by axis is F 0.423, C 0.235, P 0.146, E 0.132, and Q 0.121; only F exceeds baseline overall, while Q wins none of its four dimensions.The ordering tracks external checkability: an upstream output for F, a peer artefact for C, and no external referent for Q.Source: §E.2 · p. 44, Table 17 · p. 45

E.3 By modality: one criterion degrades from text to pixels to time and audio

• For storyboard-script text, E is +0.175 above baseline (4/4 dimensions), versus 0.190 below on video time and audio (0/11); F declines from +0.317 on text to +0.101 on pixels and +0.012 on time and audio.Q is below baseline in both modalities where defined, so agreement for an axis cannot be assumed to transfer across modalities.Source: §E.3 · pp. 44–45

E.4 By human consensus: accurate where people agree

• Across 63 dimensions, the human baseline and automated scorer correlate at PLCC 0.418 (SRCC 0.416), indicating they are accurate and inaccurate on the same dimensions.When both perform poorly, criterion clarity—not annotation budget or prompt engineering—is the stated obstacle; the human baseline is the appropriate filter for retiring dimensions.Source: §E.4 · p. 45, Figure 12 · p. 45

F Case Study

• The case study uses artefacts to expose what a single score hides, trace defect origins, break down model performance, and show a case where the automated scorer fails.The cases are ordered from hidden score details to defect provenance, model breakdown, and finally automated-scorer failure.Source: F Case Study · p. 46

F.1 Equal scores, different failure modes

• gpt-image-1.5 scored 1.33 on action interaction with 13 limb boxes but 3.33 on face identity; wan2.7-image-pro scored 1.67 on face identity with 15 face boxes but 3.0 on action interaction.The models trade which dimension they fail: the model losing identity holds action at 3.0, while the model losing action holds identity at 3.33.Source: §F.1 · pp. 46–47
• Figure 13’s four items scored 1.3–1.7 on a dimension, yet their annotated defects fell in four disjoint locations, so a composite score would group unlike failures.The boxes marked faces, limbs mid-action, people who should not be present, and hands with the wrong number of fingers.Source: Figure 13 · p. 46, §F.1 · p. 46

F.2 Cross-shot consistency is a separate capability

• Two models using the same episode and character sheets differed by 2.66 points on MI-D1 cross-shot character consistency, although every failing-row panel was defensible alone.The failure was relational: a shot was wrong relative to another shot, not necessarily incorrect in isolation.Source: Figure 14 · p. 47, §F.2 · p. 47
• The top-tier row holds its two leads across shots, while the bottom-tier row fails cross-shot consistency despite individually defensible panels.The assessment concerns whether shots remain consistent with one another, a different quantity from single-asset quality.Source: §F.2 · p. 47

F.3 Shot count decides executability

• The three-shot storyboard compressed the episode into 30 s and 10 dialogue turns, with all annotators identifying the same executability and rhythm defects.They marked multiple actions in one shot, overly complex blocking, and a missing reaction shot.Source: §F.3 · p. 47, Table 18 · p. 47
• At 3 shots, all annotators scored executability and shooting rhythm 1.00 with zero disagreement; 6 shots recovered both F dimensions, while 12 improved only executability.All three models received the same episode script and differed only at the storyboard stage.Source: Table 18 · p. 47, §F.3 · p. 47

F.4 Each input paradigm carries its own failure mode

• 1/10 failing dimensions for first/last-frame inputs versus 9/15 for grid inputs shows that difficulty depends on paradigm, not visual style.Pooling paradigms dilutes the grid collapse.Source: Figure 16 · p. 48, §F.4 · p. 48
• Grid panels omit people, while first/last-frame shots can change a person between frames because identity is not held across separately produced frames.The paired paradigms therefore have different characteristic failure modes.Source: §F.4 · p. 48
• A grid-based pixverse-c1 shot scored 1.67 on keyframe adherence, with all three annotators marking keyframe deviation across three localized intervals.The intervals were 0.29–11.04 s, 2.98–4.87 s, and 6.56–8.15 s.Source: §F.4 · p. 49
• Because the grid-paradigm prompt specifies only camera moves and dialogue, appearance and blocking depend solely on the grid image, which is harder to adhere to than a first/last-frame pair.Splitting the image into cells weakens adherence to that visual anchor.Source: §F.4 · p. 49

F.5 An upstream defect is re-expressed downstream, not repaired

• 2.42 was the corpus-wide score for cross-shot scene consistency, the lowest of the four episode-level dimensions.The reported pattern was rooms drifting before people did.Source: Figure 17 · p. 49
• The Fig. 10 comparison used four dramas, a fixed video model, and first/last-frame inputs, replacing exactly one upstream stage per run across 12 configurations.The baseline chain used the leader at every stage, scoring 3.03–3.29 across the three granularities; replacements used hy3 or mimo-v2.5-pro for storyboards and nano-banana-pro or seedream-5.0-lite for images.Source: §F.5 · p. 49
• An upstream set-dressing drift was not repaired by the video stage but was re-expressed across consecutive shots and persisted into the short drama.In the illustrated run, annotators circled the set dressing while the characters held.Source: §F.5 · p. 49, Figure 17 · p. 49

F.6 hy3 reproduces the script and does not dramatise it

• On storyboard design, hy3 scores 3.43 overall versus doubao-seed-2.1-pro at 3.60, yet leads on all script-reproduction leaves and trails on all expressiveness leaves across 533 items.Dialogue fidelity is 4.44, 0.20 behind the leader; emotional expression (3.64) and audiovisual style treatment (3.91) rank eighth of nine.Source: §F.6 · p. 49, Table 19 · p. 50
• On one identical episode script, hy3 cut 10 shots and reproduced all 20 dialogue turns verbatim, while doubao-seed-2.1-pro cut 7 shots and wrote them frame by frame.Source: §F.6 · p. 49

F.7 seedance-2.5 buys episode-level coherence with some single-clip quality

• Seedance-2.5 follows prompts more literally, faithfully executing deficiencies already present in them and contributing to the single-shot losses.The comparison uses paired items matched on drama, episode, and shot; the figure reports 39/14/10 pairs across multi-reference, animated, and matched settings.Source: §F.7 · p. 49, Figure 18 · p. 50
• Shot duration also changes from 15 s to 30 s, so each shot contains four internal cuts while its score remains a single number.Source: §F.7 · p. 49
• Figure 18 shows seedance-2.5 gaining at episode and short-drama granularity while losing on single-shot video; cross-shot style consistency rises +1.143 and audio-visual sync falls −0.658.Other timeline-continuity dimensions, including motion smoothness and camera plausibility, also worsen.Source: §F.7 · p. 49, Figure 18 · p. 50

F.8 Automated scoring drifts on dimensions with no reference to check against

• In one nano-banana-pro item, every dimension with a reference is scored correctly, but the purely perceptual dimension is overestimated by 3.67 points in the lenient direction.This item illustrates the failure mode that metric validation found over the population (Sec. 3.4, Tab. 17).Source: §F.8 · p. 50, Figure 19 · p. 50